Πλοήγηση ανά Συγγραφέας "Iosifidis, Petros-Konstantinos"

Τώρα δείχνει 1 - 1 of 1
  • Μικρογραφία εικόνας
    Τεκμήριο
    Integration of gene expressions with polymorphism data in systems biology: comparison with imaging techniques in gliomas.
    (ΕΛ.ΜΕ.ΠΑ., ΣΧΟΛΗ ΜΗΧΑΝΙΚΩΝ (ΣΜΗΧ), Τμήμα Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών, 2022-11-01) Iosifidis, Petros-Konstantinos; Ιωσηφίδης, Πέτρος-Κωνσταντίνος
    The present work outlines core aspects of machine learning in the fields of radiomics, genomics, transcriptomics and radiogenomics. More specifically, it’s attempting through the usage of multi-type data (including medical images, gene expressions, trascriptome expressions) to advance the di-agnostic power of predictive models. In the same time, it’s trying to advance the survival rate met-rics using the same type of data in order to help with cancer correlations and treatment observation and evaluation. Starting off the reader will understand core concepts of the biomedical field, the nature of the problem as well as the scope and target of this thesis. Continuing we will also give the reader the necessary computational knowledge needed to follow up with the experiments. Moving forward we perform a multi-type experiment attempting to merge radiogenomic classifiers with a better cancer survival rate. Lastly we present our results, give our outlook and discuss about the work done & problems we encountered and close off by pondering over future research. The begin of the experiments starts with a lengthy preprocessing of approximately 4000 MRI blocks of multiple modalities (FLAIR, T1, T1CE, T2) and generation of custom input objects. Through the use of a DNN, namely a 3D CNN with modified inputs, we establish cancer classifica-tion and semantic segmentation into 4 major classes(background, necrotic core/non enhancing tumor, peritumoral edema, enhancing tumor) through the training and evaluation of multiple segmentation models. Using the imaging data, we extract a plethora of imaging features that we later use in gradient boosting (XGBOOST) to approximate survival prediction from the imaging data analysis. Continu-ing with the genomic & trascriptomic data, we establish two major classes of “dead” or “alive for over 100 days” and generate classifiers based on the multi-omic profiling of our samples. Lastly we use the multi-omic data to generate powerful regressors for survival rate prediction.